Local and self-hosted AI

Desktop and self-hosted workspaces combine document chat, agents, local or cloud models, and MCP tools in an AI workspace the user can run under their control.

An Apple-Silicon-native harness with a local API, model choices, and optional macOS sandbox features makes local execution boundaries the main comparison point for Mac agents.

Local runtime controls and a packaged WebUI show a lightweight personal automation workspace with visible goals and ongoing jobs.

Container-isolated agents and messaging channels provide task separation around memory and scheduled work in a local personal workflow.

A Docker setup with local and cloud model choices, cited web search, and file uploads exposes where a self-hosted answering workspace may still mix local and external services.

Local Whisper or Parakeet transcription plus local or external summary models make the processing boundary visible for a desktop meeting workflow.

A conversational meeting-note approach adds active interaction to the comparison instead of stopping at a transcript.

Reusing local coding-agent CLI sessions with a sandboxed preview shows notes and data becoming exportable visual artifacts without a separate hosted editor.

A local runtime, SQLite-backed structured memory, permission modes, and an Ollama route separate on-machine work from optional provider and messaging connections.

A 12B multimodal checkpoint with native audio support brings richer inputs into the on-device model comparison.

At 230M parameters, with ONNX, GGUF, and MLX variants, this model puts startup footprint and runtime choice ahead of broad capability in the edge comparison.

The small hybrid design provides a middle step in on-device capacity before billion-parameter hardware needs.

With 1.5B active parameters and local formats, this model exposes the tradeoff between tool-use capacity and per-token activation at the edge.

A 1B-class model with 131K context and Ollama, LM Studio, GGUF, and MLX paths makes broad local-runtime compatibility the comparison point.

An ONNX Runtime path across browser, mobile, desktop, and edge devices shows local text-to-speech moving between hardware targets.

OpenVINO export, quantization, and compression show how Hub models can be adapted for Intel CPUs, Arc GPUs, or Core Ultra NPUs.

Apple Core AI Models provides export recipes, model assets, and Swift runtime utilities for supported models that need to run inside macOS or iOS applications.

Local detection and masking of personally identifiable text provide a privacy step to inspect and test before material enters a separate model.

An opinionated on-device Whisper CLI shows a direct local transcription step without a larger meeting or voice platform.

A local SDXL interface with inpainting, outpainting, and image prompts provides an on-device image-creation path without a hosted editor.

OCR, layout parsing, and structured Markdown and JSON outputs turn local documents into machine-usable input for RAG or agents.

Reading-order, table, layout, and math-aware analysis show what can be extracted from complex pages beyond plain text recognition.

Local GPU and remote-compatible conversion paths for PDFs and document images expose the route from source pages to Markdown before downstream AI processing.

Ollama-OCR sends images and PDFs through locally served Ollama vision models, keeping document extraction on a user-controlled model path before the text enters another workflow.

TaxHacker puts receipt and invoice extraction in a self-hostable workflow with local-model options, structured fields, exports, and manual entry for sensitive financial records.